Weather models are massive computer programs that simulate the atmosphere using the laws of physics. They take in real-time observations — from satellites, radar, weather balloons, ocean buoys, aircraft and surface stations — and use equations governing temperature, pressure, wind and moisture to predict how the atmosphere will evolve.
Because the atmosphere is chaotic, small changes in starting conditions can lead to different outcomes. That’s why forecasts sometimes shift.
Models are typically run several times per day and project conditions out hours, days, and in some cases, weeks into the future. They require tremendous supercomputer processing power capable of performing trillions to quadrillions of mathematical operations per second.
The major global models
These are the workhorses for forecasting from one day to sometimes more than 10 days into the future.
European model (ECMWF)
Operated by the European Centre for Medium-Range Weather Forecasts, based in Reading, England, this model is often referred to simply as “the European” or "the Euro/"
It consistently ranks at the top in global skill scores — statistical measures of forecast accuracy.
But it is not infallible. It sometimes strengthens storms too much or latches onto a solution too aggressively. High skill does not mean always right.
American model (GFS)
The Global Forecast System (GFS) is run by NOAA’s National Centers for Environmental Prediction in the United States.
It is freely available worldwide and updates four times daily. In recent years it has undergone major upgrades but its accuracy still trails the European.
The GFS often handles broader pattern shifts well, though it can sometimes move systems along too quicky.
Canadian model (GEM)
Run by Environment and Climate Change Canada, the Global Environmental Multiscale Model (GEM) is another global system.
It can sometimes be too cold and too aggressive with storm systems. While not typically leading global skill scores, it provides valuable alternative solutions and can occasionally outperform higher-ranked models in specific setups.
UK Met model
The UKMet global model, operated by the United Kingdom’s Met Office, is widely respected. It ranks as the second most accurate global model.
ICON (German) model
Germany’s ICON model, run by the Deutscher Wetterdienst, Germany's national weather service, is a global model with high resolution and strong physics packages. It has become more prominent in recent years and sometimes handles European-Atlantic patterns particularly well.
Short-range models
When storms are imminent, we shift to higher-resolution, short-range models. Here, high-resolution means an ability to resolve smaller atmospheric flows and properties, such as clusters of thunderstorms and terrain-induced circulations. They are especially useful once storms are within 1–2 days of arrival.
These include:
NAM (North American Model)
Run by NOAA, it focuses on North America and extends out about 84 hours. An older model, it can be helpful in winter storms but has a known bias toward overdoing precipitation in some setups.
Much to the consternation of many meteorologists, the National Weather Service announced the NAM will be taken out of commission on August 31, 2026.
RGEM
Canada’s regional model. It often handles small-scale features and winter precipitation placement well.
HRRR (High-Resolution Rapid Refresh)
Updated hourly and extending 18–48 hours out, the HRRR is invaluable for severe weather, squall lines and rapidly evolving systems. It is one of the most trusted short-term tools in thunderstorm setups.
RRFS (Rapid Refresh Forecast System)
A newer, high-resolution U.S. model that aims to eventually replace or consolidate several regional systems. It is still evolving but shows promise. The Weather Service says it will take the place of the NAM.
The rise of AI models
Over the past two years, artificial intelligence has entered the forecasting space. Unlike the models summarized above, they do not solve equations and rely on math and physics. Instead, using machine learning and neural networks, they are trained on decades of atmospheric data and patterns to predict the future.
Computationally efficient, they have shown impressive early skill, particularly in pattern recognition 3 to 5 days or more into the future. But they are still being evaluated and refined.
Forecasters most commonly review three AI models:
AIFS: The Artificial Intelligence Forecasting System, developed under the ECMWF umbrella, is referred to as the Euro AI by some.
AIGFS: The AI-enhanced counterpart to NOAA’s GFS and is sometimes called the American AI.
AI WeatherNext2: This is an AI model created by Google DeepMind and Google Research.
What are ensembles?
Because the atmosphere is chaotic, forecasters don’t rely on just one model run.
Instead, we look at ensemble systems.
Models require initial data to make predictions, and observation sensors contain numerous small errors. These small errors can grow into larger ones, reducing forecast certainty.
To account for the impact of these variations, an ensemble runs the same model many times with slightly different initial conditions. Each run or simulation is called a member.
If most members cluster around a similar outcome, confidence is higher. If they diverge wildly, uncertainty increases.
Major ensemble systems include:
- GEFS (Global Ensemble Forecast System) — tied to the GFS. It has 30 members.
- EPS (Ensemble Prediction System) — tied to the European. It has 50 members.
- AI GEFS and AI EPS — AI-enhanced ensemble counterparts
Ensembles help quantify uncertainty and show ranges of possible outcomes, rather than a single deterministic solution.
Forecasters weigh ensemble forecasts more heavily during longer range predictions, of 5 or more days into the future.
Which model is “best”?
Statistically, the European has often posted the highest global skill scores at the medium range. But that does not mean it is always right.
Forecasting is not about picking a favorite model. It’s about understanding each model’s strengths, weaknesses and biases and then weighing them against real-world observations and atmospheric reasoning.
When models agree, confidence rises. When they diverge, confidence drops.
If five global models show a similar coastal storm track five days out, forecasters feel more comfortable leaning into that solution. For example, when forecasting a coastal snowstorm, if they scatter from inland runner to offshore miss, caution is warranted.
The human element still matters
Models are tools. Meteorologists like to say they're guidance, not gospel. The British statistician George Box famously said, "all modes are wrong, some are useful."
Experienced forecasters:
- Evaluate the weather pattern.
- Analyze how ensemble members differ
- Understand seasonal biases.
- Apply meteorological reasoning.
Sometimes that means siding with the consensus. Other times it means rejecting it.
In short: the most accurate forecast comes not from one model, but from synthesizing many — guided by science and experience.
That’s why when you hear us reference “the models,” it’s rarely just one. It’s an ecosystem of guidance, each offering clues to how the atmosphere may evolve.
